training on Hacker News

90 titles from 82 authors in the last 30 days, with 86 distinct links and 10 Show HN.

Counts titles, it does not confirm facts. Open the source before concluding anything.

Projects on this term

The score comes from our model.

• Show HN: Simulusion – A PyTorch library for training and running world models · score 6.8

  Simulusion is a modular PyTorch library that unifies multiple world-model algorithms (Dreamer, PlaNet, JEPA, IRIS, DIAMOND, DiT, Genie) under a consistent API, with environment adapters and export support. It stands out for its breadth of implemented algorithms, shared trainer, and reproducible demos with checkpoints on Hugging Face, making world-model research more accessible.

• Show HN: How to Speed Up VLA Training Step by Step:3× Acceleration on Pi0.5 · score 6.3

  This is a detailed technical blog post documenting optimizations to the Pi0.5 vision-language-action model training on a framework called LoongForge, achieving a 3x speedup through techniques like pruning redundant camera slots, selective torch.compile, FlexAttention, and strategic activation recomputation. It provides deep, profiling-driven insights into VLA training bottlenecks and is valuable for ML engineers working with large multimodal models.

• Show HN: Spectra – A drop-in optimizer that cuts LLM training VRAM by 50% · score 6.0

  Spectra is a drop-in replacement for AdamW that reduces optimizer state memory by 50% by factorizing second-moment variance and filtering high-frequency noise via FFT. It claims faster training than GaLore without quantization collapse, making it appealing for memory-constrained LLM training.

• Show HN: Surprisal – a language identifier in ~250 lines of C. No ML,no training · score 5.8

  Surprisal is a language identifier that uses a hand-rolled LZ77 compressor to measure how many bits it takes to encode a query given a reference text for each language. The least surprising language wins, with no machine learning or training involved. It's an interesting demonstration of the connection between compression and prediction, implemented compactly in around 250 lines of C.

Recent titles

1. We're Training a Figma-Fluent Editing Model
   2026-11-01 · figma.com · HN discussion
2. Anthropic blocks users for anything touching training larger models
   2026-10-31 · twitter.com · HN discussion
3. How to benchmark mixed-GPU training without fooling yourself
   2026-10-31 · nodeox.com · HN discussion
4. Google Adopts New AI Training Control
   2026-10-31 · seroundtable.com · HN discussion
5. Making it easier for websites to express AI training preferences
   2026-10-31 · openai.com · HN discussion
6. Google plans to support shared AI training opt-out via robots.txt
   2026-10-30 · developers.google.com · HN discussion
7. New HN: Moju – Open-source physics auditing for CFD and SciML training data
   2026-10-30 · github.com · HN discussion
8. We're training a Figma-fluent editing model
   2026-10-30 · twitter.com · HN discussion
9. Show HN: Anime Text Datasets for LLMs Training
   2026-10-30 · github.com · HN discussion
10. AI training platform Outlier illegally surveils European platform workers
    2026-10-29 · noyb.eu · HN discussion
11. Jabhook: Harnessing computer vision for martial arts training
    2026-10-28 · jabhook.ai · HN discussion
12. Training Reve 2.0 at Scale
    2026-10-28 · app.reve.com · HN discussion
13. OpenAI no longer training AI models on live internet, citing 'significant risks'
    2026-10-28 · techcrunch.com · HN discussion
14. AI's Rogue Behavior Starts with Its Training
    2026-10-28 · nytimes.com · HN discussion
15. Ask HN: Should there be an open source license that bans model training
    2026-10-28 · HN discussion
16. Show HN: Simulusion – A PyTorch library for training and running world models
    2026-10-28 · github.com · HN discussion
17. Training a stock scoring model 800 times taught me six ways it fools itself
    2026-10-28 · stockdataanalytics.com · HN discussion
18. Show HN: Which big sites block AI training crawlers but let AI search in
    2026-10-27 · getkavel.com · HN discussion
19. Officially corrected errors persist in AI training data
    2026-10-27 · zenodo.org · HN discussion
20. Show HN: Pandora M1, a distributed training GPU monitor box
    2026-10-26 · blog.christianperone.com · HN discussion
21. Open source models are much worse at training other models
    2026-10-26 · frontierml.org · HN discussion
22. Batchplan: Plan batches of robot motions, use for execution or training data
    2026-10-24 · github.com · HN discussion
23. Amazon's Trainium Takes a Strike at the Training King
    2026-10-24 · newsletter.semianalysis.com · HN discussion
24. OpenAI disables internet access during model training
    2026-10-24 · theregister.com · HN discussion
25. OpenAI: Securing our frontier model training environment
    2026-10-23 · openai.com · HN discussion
26. Why Training AI to Have a Self Makes Governance Harder (audio)
    2026-10-23 · packetpushers.net · HN discussion
27. Show HN: How to Speed Up VLA Training Step by Step:3× Acceleration on Pi0.5
    2026-10-23 · baidu-baige.github.io · HN discussion
28. GLM 5.3 internally considers its responses influence on next model training
    2026-10-22 · lesswrong.com · HN discussion
29. Databricks launches AI Runtime for managed GPU training on enterprise data
    2026-10-22 · databricks.com · HN discussion
30. Reactive Optimizers: rewrite the optimizer's source code during training
    2026-10-21 · xingjianbai.com · HN discussion

All trends